Element-Wise Femoral Stress Prediction: A Hybrid Graph Neural Network Surrogate for Finite Element Analysis
Organizations: BCN MedTech, Department of Engineering, Pompeu Fabra University, Barcelona, Spain · 3D-Shaper Medical Barcelona, Spain · ICREA, Barcelona, Spain
Abstract
Hip fractures represent a significant clinical challenge because they cause excess disability and mortality, and have a growing incidence in aging populations, as bone fragility increases. Although finite element analysis (FEA) provides biomechanical assessment of femoral strength, its clinical applicability is mainly limited by high computational cost. Conventional machine learning (ML) approaches have been explored as surrogate models for FEA; predictions are fast but fail to capture the spatial and topological structure of finite element meshes. To address these limitations, this study proposes a hybrid Graph Neural Network (GNN)-based framework for femoral stress prediction using graph-structured finite element modeling (FEM) data. The proposed GCN-SAGE-Residual architecture combines graph convolution, neighborhood aggregation, and residual learning to model spatial dependencies and stress propagation across mesh elements. We systematically evaluated 14 models in a primary comparison, spanning classical ML methods and baseline GNN architectures, complemented by a targeted architectural ablation that isolates the contribution of each component of the proposed design. Evaluation results demonstrate that the proposed hybrid architecture achieves the strongest predictive performance among all evaluated models. The proposed hybrid model achieves the best performance with an R^2 score of 0.8022 and mean absolute error of 2.879 MPa, highlighting the importance of incorporating mesh topology in biomechanical prediction tasks. The proposed framework provides a scalable and computationally efficient alternative for FEM-based orthopedic analyses and fracture risk assessment. The code is publicly available on GitHub.
Figures & tables
| Age | Height (cm) | Weight (kg) | aBMD (g/cm 2 ) | |
| Control (n = 64) | ||||
| Fracture (n = 64) | ||||
| -value | 0.957 | 0.892 | 0.722 | 0.001 |
| Property | Value | Interpretation |
|---|---|---|
| Nodes per graph | 126,800 | Uniform mesh discretization |
| Edges per graph | 375,240 | Face-adjacency connectivity |
| Average degree | Hexahedral face-adjacency connectivity | |
| Degree range | 4–6 | Boundary (4) to interior (6) |
| Graph density | Sparse local connectivity | |
| Sparsity | Efficient sparse operations |
| Model | Type | R² | MAE (MPa) |
|---|---|---|---|
| Ridge | Linear ML | 0.1048 | 6.963 |
| Lasso | Linear ML | 0.1017 | 6.888 |
| ElasticNet | Linear ML | 0.0976 | 6.899 |
| Gradient Boosting | Ensemble ML | 0.7626 | 3.246 |
| Random Forest | Ensemble ML | 0.7617 | 3.177 |
| CatBoost | Ensemble ML | 0.752 | 3.394 |
| Model | Architecture Type | #Parameters | RMSE | MAE | R² |
|---|---|---|---|---|---|
| GCN-128 | Single (GCN) | 94406 | 8.402 | 3.321 | 0.7300 |
| GCNResidual-128 | Single (GCN+Res) | 127942 | 8.469 | 3.302 | 0.737 |
| GAT-128 | Single (GAT) | 95942 | 9.217 | 3.645 | 0.6835 |
| SAGE-128 | Single (SAGE) | 177862 | 8.55 | 3.342 | 0.7294 |
| GCN-256 | Single (GCN) | 369030 | 8.089 | 3.131 | 0.7558 |
| SAGE-256 | Single (SAGE) | 699782 | 8.101 | 3.159 | 0.7538 |
| Configuration | #Params | R 2 | MAE (MPa) | RMSE (MPa) |
|---|---|---|---|---|
| GCN-128 (single) | 94,406 | 0.7300 | 3.3210 | 8.4020 |
| SAGE-128 (single) | 177,862 | 0.7294 | 3.3421 | 8.5500 |
| GCN+Residual-128 | 127,942 | 0.7370 | 3.3021 | 8.4690 |
| SAGE+Residual-128 | 179,526 | 0.7140 | 3.5263 | 8.8733 |
| GCN+SAGE, No Residual-256 | 640,198 | 0.7507 | 3.1210 | 8.2238 |
| GCN-SAGE-Residual (proposed) | 709,318 | 0.8022 | 2.879 | 7.2190 |
| Stage | Time (per patient) | Hardware |
| FE simulation (Abaqus) | 51 min | Section 2.3 |
| Data preprocessing | 78.74 5.67 ms | Intel Core i9-13900HX (24 cores) |
| Device transfer (CPU GPU) | 4.05 1.35 ms | NVIDIA RTX 4090 Laptop GPU |
| GNN model inference | 98.45 2.06 ms | NVIDIA RTX 4090 Laptop GPU (17.2 GB VRAM) |
| Total GNN surrogate time | 181.24 ms | |
| Speed-up factor | 16,884 |
Neural Surrogate Modeling
Neural surrogate modeling uses learned models to approximate costly simulations, enabling faster prediction, design, control, and inference.
See all 472 papers